多尺度特征双线性融合的骨龄自动评估方法

Bone age automatic assessment method based on bilinear fusion of multi-scale features

针对传统人工评估骨龄方法受主观因素影响大、专业性要求高这一问题,提出一种多尺度特征双线性融合的骨龄自动评估框架。由感兴趣区域(ROI)分割网络和骨龄回归网络组成。ROI分割网络利用类激活图映射机制在弱监督条件下分割出不同尺度的手骨区域作为回归网络的输入。回归网络以EfficientNet作为特征提取器,通过嵌入双线性池化模块进行多尺度特征融合,增加性别信息以平衡不同性别手骨发育的差异,并利用分层交叉验证来降低数据集划分对实验结果的影响。最后在临床数据集和RSNA骨龄挑战数据集上评估模型,预测骨龄的平均绝对值误差分别为4.86个月和5.77个月。实验结果表明:嵌入双线性池化模块融合多尺度特征能有效提升骨龄评估精确度,可应用于临床辅助评估。

Aiming at the problem that traditional manual bone age assessment methods are greatly affected by subjective factors and professional requirements are high,an automatic bone age assessment framework based on bilinear fusion of multi-scale features is proposed.It is composed of region of interest(ROI)segmentation network and bone age regression network.The ROI segmentation network uses class activation graphs mapping mechanism to segment the hand skeletal regions of different scales as input of the regression network under weak supervision.The regression network uses EfficientNet as feature extractor,by embedding bilinear pooling module to carry out multi-scale feature fusion,adds gender information to balance the differences in hand bone development between different genders,and uses hierarchical cross-validation to reduce the impact of dataset partition on the experimental results.Finally,model assessment is carried out on the clinical dataset and RSNA bone age challenge dataset,the mean absolute error of predicting bone age is 4.86 months and 5.77 months,respectively.The experimental results show that the embedded bilinear pooling module fuses multi-scale features can effectively improve the accuracy of bone age assessment,which can be applied to clinical auxiliary assessment.